Hcl-model

Changed
- Refactor `hcl_model.utils.decayed_weights.decayed_weights` and allow `None` as input for `full_weight_obs` and `downweight_order`. See more details in the docstring.

Changed
- Rename argument `full_weight_weeks` to `full_weight_obs` in `hcl_model.transformers.truncate.TruncateTransformer`. Truncation is now calculated in terms of number of observations instead of datetime indexing to accommodate numpy arrays.

Changed
- Break down `CalendarTransformer` into `sklearn` compatible transformers: `AddAutomaticSeasonalDummies`, `AddHolidayDummies`, `AddHolidayTriangles`, and `AddPeriodicSplines`.

Added
- Trivial `inverse_transform` method that returns the input to `TruncateTransformer`, `TargetOutlierCorrectionTransformer`, and `TargetStructuralBreakCorrectionTransformer`. This addition allows using these transformers in `sklearn.compose.TransformedTargetRegressor` as `tranformer` argument for trivial inverse transform of pipeline predictions.
- Argument `weights` in `HandCraftedLinearModel`. It is a function that is applied to endogenous variable and passed to WLS for fitting.
Changed
- Expose fitted attributes with trailing underscore: `x_train_`, `y_train_`, `fit_results_`.
- Expose model parameters, e.g. `endog_transform` and `exog_transform` in `H andCraftedLinearModel`.

Changed
- If `num_steps` is `None`, get it from number of `X` observations in `hcl_model.model_base.ModelBase.predict`.

Changed
- Allow `np.ndarray` type for `TargetOutlierCorrectionTransformer`, and `TargetStructuralBreakCorrectionTransformer`.